CN111832438A - Electroencephalogram signal channel selection method and system for emotion recognition and application - Google Patents
Electroencephalogram signal channel selection method and system for emotion recognition and application Download PDFInfo
- Publication number
- CN111832438A CN111832438A CN202010593365.5A CN202010593365A CN111832438A CN 111832438 A CN111832438 A CN 111832438A CN 202010593365 A CN202010593365 A CN 202010593365A CN 111832438 A CN111832438 A CN 111832438A
- Authority
- CN
- China
- Prior art keywords
- electroencephalogram
- channel
- data
- emotion recognition
- channel selection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000008909 emotion recognition Effects 0.000 title claims abstract description 49
- 238000010187 selection method Methods 0.000 title claims abstract description 27
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 56
- 238000000034 method Methods 0.000 claims abstract description 41
- 238000007637 random forest analysis Methods 0.000 claims abstract description 32
- 238000001228 spectrum Methods 0.000 claims abstract description 22
- 238000011156 evaluation Methods 0.000 claims abstract description 16
- 230000003993 interaction Effects 0.000 claims abstract description 6
- 238000010801 machine learning Methods 0.000 claims abstract description 5
- 230000008451 emotion Effects 0.000 claims description 35
- 238000002474 experimental method Methods 0.000 claims description 27
- 238000007781 pre-processing Methods 0.000 claims description 16
- 230000037007 arousal Effects 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000000638 stimulation Effects 0.000 claims description 8
- 230000009977 dual effect Effects 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 7
- 238000013145 classification model Methods 0.000 claims description 5
- 230000008859 change Effects 0.000 claims description 4
- 238000011160 research Methods 0.000 description 11
- 238000012706 support-vector machine Methods 0.000 description 10
- 230000010354 integration Effects 0.000 description 8
- 238000000605 extraction Methods 0.000 description 7
- 238000005070 sampling Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- PIZHFBODNLEQBL-UHFFFAOYSA-N 2,2-diethoxy-1-phenylethanone Chemical compound CCOC(OCC)C(=O)C1=CC=CC=C1 PIZHFBODNLEQBL-UHFFFAOYSA-N 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 230000003595 spectral effect Effects 0.000 description 4
- 241000282414 Homo sapiens Species 0.000 description 3
- 238000010219 correlation analysis Methods 0.000 description 3
- 238000003066 decision tree Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000003491 array Methods 0.000 description 2
- 230000002349 favourable effect Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 206010048909 Boredom Diseases 0.000 description 1
- 206010062519 Poor quality sleep Diseases 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 210000003169 central nervous system Anatomy 0.000 description 1
- 238000007635 classification algorithm Methods 0.000 description 1
- 235000009508 confectionery Nutrition 0.000 description 1
- 210000000750 endocrine system Anatomy 0.000 description 1
- 238000013210 evaluation model Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000000653 nervous system Anatomy 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 230000008092 positive effect Effects 0.000 description 1
- 210000004761 scalp Anatomy 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention belongs to the technical field of machine learning and intelligent human-computer interaction, and discloses an electroencephalogram signal channel selection method, system and application for emotion recognition, wherein the method and system are used for carrying out degaussing pretreatment on electroencephalogram data; then, calculating by combining a sliding window and fast Fourier transform to obtain the power spectrum intensity of the frequency domain signal, and taking the power spectrum intensity as the electroencephalogram characteristic; respectively adopting Relieff and MIC algorithms to obtain the weight of each feature, integrating by using a wave arrival counting method to obtain the integral sum of each channel, sequentially adding feature data of a channel with a larger integral sum value, and adopting random forest classification to find out an optimal channel subset; and performing classification evaluation. The invention adopts a characteristic selection method combining the Relieff algorithm and the MIC algorithm, and takes each channel as a whole, thereby realizing the aim of greatly reducing the number of channels, improving the efficiency of the system and the real-time property of the system, and having important significance for the fields of electroencephalogram emotion recognition and intelligent human-computer interaction.
Description
Technical Field
The invention belongs to the technical field of machine learning and intelligent man-machine interaction, and particularly relates to an electroencephalogram signal channel selection method and system for emotion recognition and application.
Background
In recent years, emotion recognition has become a popular topic in the fields of emotion calculation, computational neuroscience, human-computer interaction and the like, and meanwhile, the emotion recognition has been widely applied to a plurality of fields such as medical treatment, education, games, aviation and the like. At present, many scholars select electroencephalogram signals to carry out related research on emotion recognition for the following reasons. Firstly, the electroencephalogram signal belongs to a physiological signal, and the physiological signal is directly controlled by a nervous system and an endocrine system, has spontaneity and is not influenced by subjective consciousness of human beings; the electroencephalogram signal is directly generated by the central nervous system closely related to emotion, the extraction is simple, the time resolution is high, the real-time performance is strong, the activity state of the brain can be directly reflected, and meanwhile, the emotion recognition rate is high.
The time resolution of the electroencephalogram signal is higher, but the spatial resolution of the electroencephalogram signal is lower, and in order to acquire more abundant information, more electrodes are usually placed on the scalp of a subject, and the electroencephalogram signal of multiple channels (32 or 64 or 128 channels) is generally adopted for emotion recognition research, so that a high recognition rate is expected to be achieved. However, after too many channels of electroencephalogram signals are used, equipment cost is increased and operation complexity is increased, extra electroencephalogram channels may include noise and redundant channels, recognition performance is reduced, meanwhile, too much calculation amount is caused by too many dimensions of feature data, instantaneity of an emotion recognition system is affected, and accordingly, efficiency of emotion recognition processing is reduced.
Through the above analysis, the problems and defects of the prior art are as follows: the problems of large data quantity, high cost, low data redundancy and instantaneity and the like exist in the electroencephalogram emotion recognition.
The difficulty in solving the above problems and defects is: the channel selection of the electroencephalogram signal inherently reduces the data dimension and the computational complexity, but how to reduce the number of channels as much as possible while minimizing the loss of classification accuracy is thought and balanced. In addition, there is individual difference between the test subjects, so there may be difference in channel selection results obtained for different test subjects.
The significance of solving the problems and the defects is as follows: and finding a channel subset with the maximum emotion relevance, and removing redundant and irrelevant channels, thereby reducing the dimensionality of the feature data, improving the system efficiency and improving the system identification performance. And the electroencephalogram signal channel which is irrelevant to the tested electroencephalogram signal channel, namely the public channel is selected, so that the emotion recognition system has universality.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides an electroencephalogram signal channel selection method and system for emotion recognition and application.
The electroencephalogram signal channel selection method for emotion recognition is realized in the way that the electroencephalogram signal channel selection method for emotion recognition is used for performing degating pretreatment on electroencephalogram data; calculating by combining a sliding window and fast Fourier transform to obtain the power spectrum intensity of the frequency domain signal as the electroencephalogram characteristic; then, respectively adopting Relieff and MIC algorithms to obtain the weight of each feature, integrating by using a wave arrival counting method to obtain the integral sum of each channel, sequentially adding feature data of a channel with a larger integral sum value, and adopting random forest classification to find out an optimal channel subset; and (4) carrying out classification evaluation by adopting SVM, random forest and KNN.
Further, the electroencephalogram signal channel selection method facing emotion recognition comprises the following steps:
firstly, preprocessing electroencephalogram data: after the electroencephalogram data are obtained, firstly, a base value removing operation is carried out, the electroencephalogram data in a calm state of the first few seconds are recorded as base value data, the electroencephalogram data under video stimulation are recorded as original data, and the difference value of the two data reflects the relative change of the electroencephalogram when a person generates emotion;
secondly, extracting electroencephalogram features: setting a sliding window, converting input time domain data into a frequency domain by adopting fast Fourier transform aiming at each window, and solving the power spectrum intensity of a frequency domain signal to be used as an electroencephalogram characteristic;
thirdly, selecting an electroencephalogram channel: according to the correlation of the features and the categories, respectively adopting a Relieff algorithm and an MIC algorithm to obtain the weight of each feature, integrating the weight of the included features of each channel by a wave arrival counting method, performing descending order arrangement on the channels according to integral sum, gradually increasing the number of the channels, classifying by using a random forest, and searching for the optimal channel subset according to the recorded result.
Further, the specific implementation process of acquiring the experimental input data is as follows: using 1s as a division criterion, XiBase line data representing the i (i ═ 1,2.., m) th second, Base value, RawjRepresents the raw data under j (j ═ 1,2.., n) th second of video stimulation, InputjI.e. the j-th second of the experiment, the related calculation formula is shown as the following formula:
Inputj＝Rawj-Base。
further, the specific implementation process of the integrated channel integral sum is as follows: after weights of n-dimensional electroencephalogram features on g frequency bands are respectively obtained by adopting a Relieff algorithm and an MIC algorithm, weights of all features contained on one channel are integrated by adopting a wave arrival counting method, namely 2g groups are total, the features of each group are arranged in a descending order, integrals n-i are given to the features according to the sequence i of each feature, and the integrals of all the features of the 2g groups on each channel are summed to obtain the integral sum of the channels.
Further, the specific implementation process of finding the optimal channel subset is as follows: performing descending order arrangement on the c channels according to the size of the integral sum, and starting an iterative experiment; and initializing the channel number i to be 0, if i is less than the total channel number c, adding characteristic data of a channel which is ranked in front, classifying the channel by adopting a random forest, recording the index number of the channel and the corresponding recognition rate of the index number, drawing a curve of which the classification accuracy rate changes along with the channel number, finding out the optimal point with less channel number and low precision loss according to the curve, and corresponding to the optimal channel subset.
Further, the emotion classification model includes three types: the emotion classification of the Valence dimension is two, the emotion classification of the Arousal dimension is two, and the emotion classification of the Valence-Arousal dual dimension is four.
Another object of the present invention is to provide an electroencephalogram signal channel selection system for emotion recognition, which operates the electroencephalogram signal channel selection method for emotion recognition, and the electroencephalogram signal channel selection system for emotion recognition comprises:
the electroencephalogram data preprocessing module is used for performing degaussing preprocessing on the electroencephalogram data;
the frequency domain signal power spectrum intensity acquisition module is used for calculating and obtaining the power spectrum intensity of the frequency domain signal by combining a sliding window and fast Fourier transform and taking the power spectrum intensity as the electroencephalogram characteristic;
the optimal channel subset acquisition module is used for solving the weight of each feature by adopting a Relieff algorithm and an MIC algorithm, integrating by using a wave arrival counting method to obtain the integral sum of each channel, sequentially adding feature data of a channel with a larger integral sum value, and classifying by adopting a random forest and finding out an optimal channel subset;
and the classification evaluation module is used for performing classification evaluation by adopting an SVM, a random forest and a KNN.
The invention also aims to provide a machine learning system which is loaded with the electroencephalogram signal channel selection system facing emotion recognition.
The invention also aims to provide an intelligent human-computer interaction system which carries the electroencephalogram signal channel selection system facing emotion recognition.
The invention also aims to provide an emotion recognition control system, which carries the electroencephalogram signal channel selection system facing emotion recognition.
By combining all the technical schemes, the invention has the advantages and positive effects that: the method is based on a DEAP data set to carry out experiments, weights of electroencephalogram characteristics are respectively obtained through Relieff and MIC algorithms, then the sum of integrals of channels is obtained through integration of a wave arrival counting method, and a classifier algorithm is adopted to carry out evaluation and verification so as to realize electroencephalogram channel selection.
Aiming at the problems of large data volume, high cost, low data redundancy and instantaneity and the like in the conventional electroencephalogram emotion recognition, the invention provides a novel electroencephalogram signal channel selection method facing emotion recognition, and aims to find a channel subset with the maximum emotion correlation and remove redundant and irrelevant channels, so that the dimensionality of characteristic data is reduced, the system efficiency is improved, and the system recognition performance is improved.
According to the method, the weights of electroencephalogram characteristics and categories are calculated according to the correlation of the characteristics and the categories by combining the Relieff algorithm and the MIC algorithm, then the integral sum of electroencephalogram channels is obtained by adopting the idea integration of the wave arrival counting method, which is the comprehensive embodiment of all characteristic weights on one channel, and compared with the method adopting a single characteristic selection algorithm, the method can find out the channel subset with strong emotion correlation, and has certain generalization.
The method can greatly reduce the number of channels under the condition of slight loss of precision, thereby solving the problems of large electroencephalogram data volume, redundancy, low system real-time property and the like in the current emotion recognition field. The number of channels can be reduced from 32 to 5 in both the value dimension and the Arousal dimension, the corresponding recognition rate can reach 90.07% and 90.47% at most, and the difference between the result and the result of the whole channel is about 3% and 2%; the highest 32 channels can be reduced to 10 channels in the Valence-Arousal four-classification model, the recognition rate can reach 89.11 percent, and the difference between the recognition rate and the full channel is about 1 percent.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments of the present application will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present application, and it is obvious for those skilled in the art that other drawings can be obtained from the drawings without creative efforts.
FIG. 1 is a flowchart of an emotion recognition-oriented electroencephalogram signal channel selection method provided by the embodiment of the invention.
FIG. 2 is a schematic structural diagram of an electroencephalogram signal channel selection system for emotion recognition provided by an embodiment of the present invention;
in fig. 2: 1. the electroencephalogram data preprocessing module; 2. a frequency domain signal power spectrum intensity obtaining module; 3. and an optimal channel subset acquisition module.
FIG. 3 is a flowchart of an implementation of an emotion recognition-oriented electroencephalogram signal channel selection method provided by an embodiment of the present invention.
Fig. 4 is a schematic diagram of a channel selection result for emotion two classification in the value dimension according to an embodiment of the present invention.
Fig. 5 is a schematic diagram of a channel selection result for emotion two classification in the Arousal dimension according to an embodiment of the present invention.
Fig. 6 is a schematic diagram of a channel selection result for performing emotion four classification on the Valence-Arousal dual dimension according to the embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Aiming at the problems in the prior art, the invention provides an electroencephalogram signal channel selection method and system for emotion recognition and application thereof, and the invention is described in detail below with reference to the accompanying drawings.
As shown in FIG. 1, the electroencephalogram signal channel selection method facing emotion recognition provided by the invention comprises the following steps:
s101: preprocessing electroencephalogram data: after obtaining the electroencephalogram data, for each experiment, the electroencephalogram data in a calm state of m seconds is recorded as a base value, the data of the subsequent n seconds is the electroencephalogram data of the person after being stimulated by the video, the data is recorded as original data, and the input data of the experiment is obtained by subtracting the base value from the original data.
S102: extraction of electroencephalogram features: setting a sliding window with the window size of l seconds and the step length of s seconds, respectively converting each window of input data into g frequency bands of a frequency domain through fast Fourier transform on c channels, carrying out weighted summation on the amplitude of an amplitude spectrum to obtain the power spectrum intensity, and taking the power spectrum intensity as electroencephalogram characteristics.
S103: selecting an electroencephalogram channel: respectively adopting a Relieff algorithm and an MIC algorithm to obtain the weight of each feature, integrating the weight of the features contained in each channel by a wave arrival counting method, arranging c channels in a descending order according to the integral sum, then gradually increasing the number of the channels, simultaneously extracting corresponding feature data, classifying by adopting a random forest, recording the channel subsets and the corresponding identification accuracy rate, and searching the optimal channel subset according to the recorded result.
The electroencephalogram signal channel selection method facing emotion recognition provided by the invention can be implemented by adopting other steps by persons skilled in the art, and the electroencephalogram signal channel selection method facing emotion recognition provided by the invention in fig. 1 is only a specific embodiment.
The invention regards the channel as a whole to carry out channel selection research, and on the basis, two schemes are to firstly carry out characteristic selection and then select the channel containing the characteristics, which can be used as the extension of the invention. The first scheme is that g frequency bands are regarded as g groups after weights of 160 features are obtained, the weights obtained by two algorithms in the groups are integrated by adopting a wave arrival counting method, and then n features in the g groups are respectively arranged from high to low according to the integral sum of the n features; each time, selecting one feature with higher integral from the g groups, then utilizing random forests to carry out classification evaluation on corresponding feature data, recording results, drawing to find out the optimal feature subset, and then analyzing to find out the channel subset containing the features. The second scheme is that a weight value obtained by integrating 160 features on two algorithms is calculated by using a wave arrival counting method, each feature obtains a corresponding integral sum, and then the 160 features are clustered and grouped by using k-means to search the association degree between the features and whether the features have the same characteristic or not. And aiming at the corresponding features of each group, performing feature descending arrangement in the group according to the integral sum, then selecting the features with higher weight value one by one from each group, performing classification evaluation by utilizing a random forest, searching the optimal feature subset, and further finding out the channels containing the features.
As shown in fig. 2, the electroencephalogram signal channel selection system for emotion recognition provided by the present invention includes:
the electroencephalogram data preprocessing module 1 is used for preprocessing the electroencephalogram data by removing the basis value.
And the frequency domain signal power spectrum intensity acquisition module 2 is used for calculating and obtaining the power spectrum intensity of the frequency domain signal by combining a sliding window and fast Fourier transform, and the power spectrum intensity is used as the electroencephalogram characteristic.
And the optimal channel subset acquisition module 3 is used for solving the weight of each feature by adopting a Relieff algorithm and an MIC algorithm, integrating by using a wave arrival counting method to obtain the integral sum of each channel, sequentially adding feature data of a channel with a larger integral sum value, and classifying by adopting a random forest and finding out the optimal channel subset.
The technical solution of the present invention is further described below with reference to the accompanying drawings.
The invention aims to reduce the number of channels on the premise of less loss of identification precision. Firstly, extracting the characteristics of electroencephalogram data by utilizing fast Fourier transform, then calculating the weight of the characteristics by combining a Relieff algorithm and an MIC algorithm, integrating the characteristics by adopting a DOA counting method to obtain the integral sum value of channels, meanwhile, carrying out classification evaluation by combining a random forest, and finally verifying the selected electroencephalogram channel subset by adopting a support vector machine, the random forest and a K nearest neighbor algorithm.
Fast Fourier Transform (FFT) is a fast algorithm obtained by improving an algorithm of discrete fourier transform according to characteristics of odd, even, imaginary, real, etc. of the discrete fourier transform, which can save a large amount of calculation cost to obtain frequency domain characteristics of signals, and thus, in practical research, FFT is generally adopted to realize efficient calculation.
The Relieff algorithm is an extension of the Relief algorithm and can process multi-classification problems, both the two algorithms belong to a feature weight algorithm, and the main idea is to endow corresponding weights of different features according to the relevance of each feature and a class, wherein the relevance is measured by the distinguishing capability of the features on different types of data. The larger the weight, the more favorable the representation of the feature for classification; conversely, a smaller weight indicates that the feature has a negative impact on the classification.
The nod counting method (BordaCount) is a ranking voting algorithm based on integration system proposed by Jean-Charles Borda in 1770, and its main idea is: and according to the sorting position of each candidate item in different groups, giving corresponding points to the candidate items, and summarizing the points obtained by each candidate item from different groups to show the position of the candidate item in the whole hierarchy.
A Support Vector Machine (SVM) is a linear classifier, and aims to find an optimal hyperplane capable of maximizing the spatial interval of different classes of feature data. Random forests (RF for short) are a combination classifier algorithm, which is a set of multiple decision trees, for each test data input into a random forest, each decision tree learns the data and selects an optimal classification result, and finally, the most predicted class of the decision trees is used as a final classification result. The K-nearest neighbors (KNN) is a non-parametric statistical method for classification and regression, which finds K training samples nearest to test sample data according to the "distance" between the test sample data and known training set samples, and then, according to the voting principle, the most appeared classes are the classification results of the test samples. All three classifiers (SVM, RF and KNN) belong to supervised learning methods.
As shown in FIG. 3, the electroencephalogram signal channel selection method facing emotion recognition provided by the invention comprises the following steps:
the method comprises the following steps: preprocessing electroencephalogram data:
the invention is based on a data set of a DEAP preprocessing version, the sampling rate of the data set is 128Hz, the frequency range is 4-45 Hz, each experiment consists of a baseline value of 3s and data under video stimulation of 60s, and physiological signals of 40 channels of a subject are recorded (the first 32 are related to electroencephalogram, and the last 8 are related to peripheral physiological signals). A total of 32 subjects were involved, each person watched 40 videos, i.e. 40 experiments. The data of each person is recorded in a file, the content comprises two parts of data and a label, wherein the data is sampling data (dimension: 40 × 8064) of the electroencephalogram signal, the label is a scoring result (dimension: 40 × 4, numerical range: 1-9 continuous values) of a self-evaluation model, and the scoring is based on four dimensions: valency pleasure, Arousal, Dominance, and Liking joy.
The blank base line of 3s before each experiment can be regarded as the electroencephalogram data of the human under the calm state and recorded as the base value, and the data of the subsequent 60s is the electroencephalogram data of the human after being stimulated by the video and recorded as the original data. In order to solve individual difference and reduce the influence of a baseline signal on emotion recognition, a lot of researches directly abandon the data of the first 3s, but considering that electroencephalogram signals can be generated no matter in a calm state or in an emotion generation state, the difference value of the electroencephalogram signals can reflect the relative change of the electroencephalogram in the emotion generation state, and the electroencephalogram characteristics of the emotion can be reflected, so that the input data of the experiment is obtained by subtracting a base value from the original data, the input data is not absolute data of the electroencephalogram any more, but is electroencephalogram fluctuation data caused by the emotion generation in the calm state.
Taking an experiment as an example, 1s is taken as a division standard, and X is assumediBase line data representing i (i-1, 2,3) th second, Base value, RawjRepresents the raw data under j (j ═ 1,2.., 60) th second of video stimulation, InputjI.e. the j-th second of the experiment, the related calculation formula is shown as the following formula:
Inputj＝Rawj-Base；
step two: extraction of electroencephalogram features:
considering that the duration of the emotion is relatively short, there are studies showing that the size of the time window is optimal between 1-2s, and therefore a sliding window with a window size of 2s (256 data points at a sampling rate of 128 Hz) and a step size of 0.5s (64 data points) is provided, with a partial overlap between adjacent windows. Then, the signals are transformed to 5 frequency bands of the frequency domain by FFT on 32 channels for each window, so that 160-dimensional electroencephalogram characteristics can be obtained for each sample, and the characteristics refer to the power spectrum intensity of the frequency domain signals, as follows.
The band is set as [4,8,12,16,25,45], that is, according to the difference of frequency ranges, the electroencephalogram signals can be divided into 5 frequency bands (θ, α, low β, high β, γ), as shown in table 1.
TABLE 1 frequency band division of electroencephalogram signals
| Theta | Alpha | LowBeta | HighBeta | Gamma |
| 4～8Hz | 8～12Hz | 12～16Hz | 16～ |
25～45Hz |
Suppose there is a time series of data x ═ x1,x2,...,xn]Then its corresponding FFT result is [ X ]1,X2,...,Xn]. Then calculating the power spectrum intensity, namely carrying out weighted summation on the amplitudes of the frequency domain signal amplitude spectrum, and setting fsIs the sampling rate, N is the length of the FFT, i.e. the length of the time sequence x, then the power spectral strength PSI of the frequency band kkIs defined as follows:
step three: selecting an electroencephalogram channel:
(1) relieff algorithm
Record training set data as DFThe total number of samples is m, and the iteration number of the algorithm is also m. Each time at DFRandomly selecting a sample R, and then finding out the sample set of the same class as RThe k samples closest to R are denoted as NearMisses, and then the k samples closest to R are found from the sample set of each category different from R, and denoted as NeatHits, and the weight update formula of each feature is shown as the following formula:
wherein R isiRepresents the ith sample, flIs RiThe first feature of (a), W (f)l) Then represents the feature flThe initial weight of the feature is 0. Hj(i ═ 1,2.. k) and RiHomogeneous ith k neighbor sample, Mj(C) (i ═ 1,2.. k) and RiThe heterogeneous jth k neighbor sample. P (C) is the prior probability of class C, i.e. the proportion of samples of class C to the total number of samples.
diff(fl,R1,R2) Representative are two samples R1And R2In the first feature flThe distance of (d) is as shown in the following formula. If R isiAnd k homogeneous neighbors HjThe sum of the distances over a feature is less than RiAnd all heterogeneous k heterogeneous neighbors Mj(C) The sum of the distances on a certain feature shows that the feature has good distinguishing capability on the same type and the different types, and then the weight of the feature is increased; on the contrary, it indicates that the feature is not favorable for classification, and its weight is reduced:
(2) MIC algorithm
The correlation analysis is a method for measuring the degree of correlation between two or more variables, and the MIC algorithm is one of correlation analysis methods, which is calculated based on mutual information (MutualInformation) and trellis division. Suppose there are two variables X ═ XiN, i ═ 1,2,. n } and Y ═ YiI 1,2.. n }, then the mutual information in the information theory can be used to measure the degree of correlation between these two variables. P (X, y) denotes X andy, and p (X) and p (Y) represent the edge probability densities of X and Y, respectively, then the mutual information is defined as follows:
let D be a set of finite ordered pairs, D { (x)i,yi) If the division G divides the X field into a parts a and the Y field into b parts, the division D is a grid of a × b. Mutual information is respectively calculated in the divided grids, so that the mutual information of the division G is equal to the maximum mutual information under the division, and the calculation formula is shown as the following formula:
MI*(D,a,b)＝maxMI(D|G)；
where D | G represents division G for data D, while MIC represents the quality of the grid by the value of mutual information, it is not only simple to calculate the size of the mutual information, but also to normalize the maximum mutual information value obtained by data D under different divisions, and then obtain the feature matrix M (D)a,b。
Feature matrix M (D)a,bIs defined as follows:
the maximum information coefficient is defined as follows:
b (n) is the upper limit of the partition grid a b, and the author gives when b (n) is n0.6The effect is the best.
(3) The concrete steps
After feature extraction, each sample has 160-dimensional features, for feature data of 5 frequency bands, a Relieff algorithm and a correlation analysis algorithm MIC are respectively used for solving the weight of 32 features on each frequency band, meanwhile, in order to find out common factors, in the view of the whole subsequent research, the corresponding feature weights of 32 persons are respectively solved, the data ranges of the weights obtained by the same algorithm are almost the same, so that the corresponding feature weights obtained by calculating the Relieff and the MIC of the 32 persons are respectively accumulated, and public channel selection is realized.
On the other hand, the difference between the feature weights obtained by the Relieff algorithm and the MIC algorithm is large, the former is 0.0005-0.0040, and the latter is 0.07-0.10, so that the feature weights calculated by the two algorithms are not directly added, and the weights of all the features contained in each channel are integrated by adopting a wave arrival counting method. There are five frequency bands and two algorithms, so there are 10 groups of individuals, according to the weight value descending ranking result of 32 features in each group, the feature ranked at the ith position can get the integral 32-i, the integral sum obtained by each channel is the basis for ranking in the whole, this is the evaluation standard for selecting the channel finally.
And arranging 32 channels in a descending order according to the integral sum, sequentially increasing the number of the channels, extracting corresponding characteristic data, then performing classification evaluation by adopting a random forest, recording the channel subsets and the corresponding classification accuracy, and drawing an accuracy change curve so as to find out the channel subsets with the smallest number of channels and stronger classification performance as far as possible.
Step four: and (3) classifier evaluation:
after the best subset of channels is selected, evaluation verification is also required on various classifiers. The classifier adopted by the invention comprises a support vector machine, a random forest and KNN, which are common classification algorithms for electroencephalogram emotion recognition research.
At present, most researchers respectively carry out sentiment classification research of two classifications aiming at Valence and Arousal dimensions, and a few researches are four classification researches based on Valence-Arousal. The criteria for the second class are bounded by 5, with values greater than 5 being considered positive classes and values less than 5 being considered negative classes. The four classification models jointly determine the classification according to the label values of Valence and Arousal, and the classification standard is shown in Table 2.
TABLE 2 Classification criteria
The invention will be described in more detail by means of the following experimental examples, which are intended for the purpose of illustration only and are not intended to limit the scope of application of the invention.
The technical effects of the present invention will be described in detail with reference to experiments.
Experiment 1: method for carrying out emotion two classification on Valence dimension by utilizing invention
In order to establish a binary model in the value dimension, the data adopted is the data only containing the value dimension in the tag array of the DEAP data set. Valence, also called Valence/pleasure, is used to measure the pleasure of emotion, such as hate and like.
Data pre-processing is performed, after baseline data of the first 3s is subtracted from data under 60s video stimulation, the data length of each experiment is changed from 63s to 60s, namely the sampling point is reduced from 8064 to 7680, and the data of each person comprises two parts: sample data (40 x 7680) and label (40 x 1).
And then, performing feature extraction, wherein the length of the preprocessed electroencephalogram time domain data is 60s, the size of a sliding window is 2s, and the step length is 0.5s, so that 116 windows, namely 116 samples, can be obtained in one experiment. By performing a fast fourier transform on these windows and then characterizing the power spectral densities over 5 frequency bands, 160 features were obtained in one experiment, with a total number of samples 4640 for one person and 148480 for the entire data set.
The correspondence between the 32 channel index numbers and the channel numbers related to the electroencephalogram is shown in table 3.
TABLE 332 channels associated with the electroencephalogram
Channel selection is then performed, and the 160 features for each person are weighted using the ReliefF and MIC algorithms, respectively, and then the corresponding feature weights for 32 persons are summed. Then, a wave arrival counting method is used to integrate the weights of the 5 features included in each channel. Because there are two algorithms and five frequency bands, there are 10 groups of individuals, and the number of available integral values arranged at the ith position is 32-i according to the descending order of the weight values of the 32 features in each group, so the sum of the integral obtained by each channel is the basis for the ordering in the whole. Then, 32 channels are arranged in a descending order according to the size of the integral sum, 5 features on one channel are sequentially added, random forests are adopted for classification, and fig. 4 is a curve that the classification accuracy of the random forests on the Valence dimension changes along with the increase of the number of the channels. Three points are marked on the graph: (5, 87.73%), (29, 91.39%) and (32, 91.15%), which respectively represent the classification accuracy rates corresponding to the channel subsets of the optimal point, the highest point and the full channel, and since the highest point and the full channel have little difference and the channel selection effect is not obvious, the subsequent classification research only focuses on the optimal point and the full channel, and the channel index numbers corresponding to the optimal point are [0,4,7,25,29], and the channel numbers corresponding to the optimal point are [ Fp1, FC5, T7, T8, P8] as known from table look-up 3.
And finally, evaluating the selected channel subset by using a classifier, wherein the classifier comprises SVM, random forest and KNN. The data sets not subjected to channel selection (i.e., full channels) are classified first, and are used as reference experiments for comparison, and then classification evaluation is performed on feature data corresponding to the optimal points. As shown in table 4, it can be seen that the recognition rate of the sweet spot is reduced by 5.4%, 2.98% and 2.84% respectively compared with that of the full channel under three classifiers, but the number of channels is reduced from 32 to 5 by about 84%, which provides a possibility for fast emotion analysis with few channels.
TABLE 4 Classification results of three classifiers on the Valence dimension
Experiment 2: emotion two classification on Arousal dimension by using the invention
In order to establish a binary model in the Arousal dimension, the data adopted is the data only containing the Arousal dimension in the tag array of the DEAP data set. Arousal also calls wakefulness, which is used to measure the intensity of emotions, such as boredom and excitement.
Data pre-processing is performed, 1s is taken as a division standard, data under 60s of video stimulation is divided into 60 data of 1s, and after the average value of the first 3s of baseline data is sequentially subtracted, the data of each subject comprises sampling data (40 × 7680) and a label (40 × 1).
And then, performing feature extraction, and after the sliding window processing with the size of 2s and the step length of 0.5s, obtaining 116 samples in one experiment. Then, a fast fourier transform is performed on each of these windows to obtain a frequency domain signal, and the power spectral densities in 5 frequency bands are used as features, i.e., one person has 4640 total samples, the whole data set has 148480 total samples, and the feature dimension is 160 dimensions.
And then, channel selection is carried out, the feature weight of each person is calculated, namely the 160-dimensional feature weight is calculated through a Relieff algorithm and an MIC algorithm, and then the corresponding feature weights of 32 persons are summed. The integration of 32 channels is then performed using a wave-arrival counting method. Then, the integral sums of the 32 channels are sorted in a descending order, then, 5 features on one channel are gradually added, and meanwhile, random forests are adopted for classification, and fig. 5 is a curve that the classification accuracy of the random forests on the Arousal dimension changes along with the increase of the number of the channels. Three points marked on the curve represent the classification accuracy corresponding to the optimal point, the highest point and the channel subset of the full channel respectively: (6, 88.40%), (29, 91.27%) and (32, 91.07%), the subsequent classification studies only focus on the optimal point and the full channel, the index number of the channel corresponding to the optimal point in the Arousal dimension is [0,3,4,7,21,25], and the channel number corresponding to the optimal point is [ Fp1, F7, FC5, T7, FC6, T8] through table look-up 3.
And finally, evaluating the channel subset data corresponding to the optimal point by using a classifier (SVM, random forest and KNN). As shown in table 5, it can be seen that the recognition rate of the optimal point is reduced by 2.48%, 2.40% and 2.40% respectively compared with that of the full channel under three classifiers, but the number of channels is reduced from 32 to 6, which is reduced by about 81%, and the channel selection effect is more obvious as in experiment 1.
TABLE 5 results of the classification of three classifiers on the Arousal dimension
Experiment 3: the method is utilized to carry out emotion four-classification on Valence-Arousal dual dimensions
In order to establish a four-classification model on the Valence-Arousal dual dimension, the data adopted is the data containing the Valence dimension and the Arousal dimension in the tag array of the DEAP data set.
Data preprocessing is carried out, and the data of each participant comprises two parts, namely sampling data (40 x 7680) and a label (40 x 2); then, feature extraction is carried out, fast Fourier transform is respectively carried out on sliding windows, the power spectral density of a frequency domain is taken as a feature, and 116 samples with 160-dimensional features can be obtained through one experiment. Then, channel selection is carried out, a weight of 160-dimensional features is calculated through a Relieff algorithm and an MIC algorithm, then, an integration sum of 32 channels is obtained through integration by adopting a wave arrival counting method, descending order arrangement is carried out according to the integration sum of the 32 channels, iteration is carried out in sequence, 5 features on the channel with a larger integration sum are added each time, meanwhile, random forests are adopted for classification, and FIG. 6 is a curve that the classification accuracy of the random forests on the Valence-Arousal double-dimension changes along with the increase of the number of the channels. Three points are marked on the curve: (10, 86.16%), (29, 87.83%) and (32, 87.69%), which respectively represent the classification accuracy rates corresponding to the channel subsets of the optimal point, the highest point and the full channel, the subsequent classification identification only concerns the optimal point and the full channel, the channel index number corresponding to the optimal point in the Valence-aroma dual dimension is [0,3,4,6,7,8,21,25,29,31], and the lookup table 3 indicates that the channel number corresponding to the optimal point is [ Fp1, F7, FC5, C3, T7, CP5, FC6, T8, P8, O2 ].
And finally, evaluating the channel subset data corresponding to the optimal point by using a classifier (SVM, random forest and KNN). As shown in table 6, it can be seen that the recognition rate of the optimal point is reduced by 2.68%, 1.40% and 1.16% respectively under three classifiers compared with the full channel, but the number of channels is reduced from 32 to 10, which is reduced by about 69%, although the number of channels is more than that of experiments 1 and 2, because the quadrant is more complicated than the binary classification, and the recognition accuracy loss is less than that of experiments 1 and 2, the effect of channel selection of experiment 3 is more obvious.
TABLE 6 Classification results of three classifiers in the valency-Arousal dual dimension
It should be noted that the embodiments of the present invention can be realized by hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or specially designed hardware. Those skilled in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such code being provided on a carrier medium such as a disk, CD-or DVD-ROM, programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier, for example. The apparatus and its modules of the present invention may be implemented by hardware circuits such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of hardware circuits and software, e.g., firmware.
The above description is only for the purpose of illustrating the present invention and the appended claims are not to be construed as limiting the scope of the invention, which is intended to cover all modifications, equivalents and improvements that are within the spirit and scope of the invention as defined by the appended claims.
Claims (10)
1. The electroencephalogram signal channel selection method for emotion recognition is characterized in that the electroencephalogram signal channel selection method for emotion recognition is used for preprocessing the basis value of electroencephalogram data; calculating by combining a sliding window and fast Fourier transform to obtain the power spectrum intensity of the frequency domain signal as the electroencephalogram characteristic; then, respectively adopting Relieff and MIC algorithms to obtain the weight of each feature, integrating by using a wave arrival counting method to obtain the integral sum of each channel, sequentially adding feature data of a channel with a larger integral sum value, and adopting random forest classification to find out an optimal channel subset; and (4) carrying out classification evaluation by adopting SVM, random forest and KNN.
2. The electroencephalogram signal channel selection method for emotion recognition as recited in claim 1, wherein the electroencephalogram signal channel selection method for emotion recognition comprises the following steps:
firstly, preprocessing electroencephalogram data: after the electroencephalogram data are obtained, firstly, a base value removing operation is carried out, the electroencephalogram data in a calm state of the first few seconds are recorded as base value data, the electroencephalogram data under video stimulation are recorded as original data, and the difference value of the base value data and the original data reflects the relative change of the electroencephalogram when a person generates emotion;
secondly, extracting electroencephalogram features: setting a sliding window, converting input time domain data into a frequency domain by adopting fast Fourier transform aiming at each window, and solving the power spectrum intensity of a frequency domain signal to be used as an electroencephalogram characteristic;
thirdly, selecting an electroencephalogram channel: according to the correlation of the features and the categories, respectively adopting a Relieff algorithm and an MIC algorithm to obtain the weight of each feature, integrating the weight of the included features of each channel by a wave arrival counting method, performing descending order arrangement on the channels according to integral sum, gradually increasing the number of the channels, classifying by using a random forest, and searching for the optimal channel subset according to the recorded result.
3. The emotion recognition-oriented electroencephalogram signal channel selection method as recited in claim 2, wherein the specific implementation process for acquiring experimental input data is as follows: using 1s as a division criterion, XiBase line data representing the i (i ═ 1,2.., m) th second, Base value, RawjRepresents the raw data under j (j ═ 1,2.., 60) th second of video stimulation, InputjI.e. the j-th second of the experiment, the related calculation formula is shown as the following formula:
Inputj＝Rawj-Base。
4. the emotion recognition-oriented electroencephalogram signal channel selection method as recited in claim 2, wherein the specific implementation process of integrating the channel integral sum is as follows: after weights of n-dimensional electroencephalogram features on g frequency bands are respectively obtained by adopting a Relieff algorithm and an MIC algorithm, weights of all features contained on one channel are integrated by adopting a wave arrival counting method, namely 2g groups are total, the features of each group are arranged in a descending order, integrals n-i are given to the features according to the sequence i of each feature, and the integrals of all the features of the 2g groups on each channel are summed to obtain the integral sum of the channels.
5. The emotion recognition-oriented electroencephalogram signal channel selection method as recited in claim 2, wherein the specific implementation process for finding the optimal channel subset is as follows: performing descending order arrangement on the c channels according to the size of the integral sum, and starting an iterative experiment; and initializing the channel number i to be 0, if i is less than the total channel number c, adding characteristic data of a channel which is ranked in front, classifying the channel by adopting a random forest, recording the index number of the channel and the corresponding recognition rate of the index number, drawing a curve of which the classification accuracy rate changes along with the channel number, finding out the optimal point with less channel number and low precision loss according to the curve, and corresponding to the optimal channel subset.
6. The emotion recognition-oriented electroencephalogram signal channel selection method, as recited in claim 3, wherein the emotion classification model comprises three types: the emotion classification of the Valence dimension is two, the emotion classification of the Arousal dimension is two, and the emotion classification of the Valence-Arousal dual dimension is four.
7. An electroencephalogram signal channel selection system for emotion recognition, which runs the electroencephalogram signal channel selection method for emotion recognition according to any one of claims 1 to 6, and is characterized in that the electroencephalogram signal channel selection system for emotion recognition comprises:
the electroencephalogram data preprocessing module is used for performing degaussing preprocessing on the electroencephalogram data;
the frequency domain signal power spectrum intensity acquisition module is used for calculating and obtaining the power spectrum intensity of the frequency domain signal by combining a sliding window and fast Fourier transform and taking the power spectrum intensity as the electroencephalogram characteristic;
the optimal channel subset acquisition module is used for solving the weight of each feature by adopting a Relieff algorithm and an MIC algorithm, integrating by using a wave arrival counting method to obtain the integral sum of each channel, sequentially adding feature data of a channel with a larger integral sum value, and classifying by adopting a random forest and finding out an optimal channel subset;
and the classification evaluation module is used for performing classification evaluation by adopting an SVM, a random forest and a KNN.
8. A machine learning system, characterized in that the machine learning system is equipped with the electroencephalogram signal channel selection system for emotion recognition according to claim 7.
9. An intelligent human-computer interaction system, characterized in that the system carries the electroencephalogram signal channel selection system for emotion recognition according to claim 7.
10. An emotion recognition control system, characterized in that, the emotion recognition control system carries the electroencephalogram signal channel selection system for emotion recognition as claimed in claim 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010593365.5A CN111832438B (en) | 2020-06-27 | 2020-06-27 | Emotion recognition-oriented electroencephalogram signal channel selection method, system and application |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010593365.5A CN111832438B (en) | 2020-06-27 | 2020-06-27 | Emotion recognition-oriented electroencephalogram signal channel selection method, system and application |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111832438A true CN111832438A (en) | 2020-10-27 |
| CN111832438B CN111832438B (en) | 2024-02-06 |
Family
ID=72899447
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010593365.5A Active CN111832438B (en) | 2020-06-27 | 2020-06-27 | Emotion recognition-oriented electroencephalogram signal channel selection method, system and application |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111832438B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112257658A (en) * | 2020-11-11 | 2021-01-22 | 微医云(杭州)控股有限公司 | Electroencephalogram signal processing method and device, electronic equipment and storage medium |
| CN112932502A (en) * | 2021-02-02 | 2021-06-11 | 杭州电子科技大学 | Electroencephalogram emotion recognition method combining mutual information channel selection and hybrid neural network |
| CN113397567A (en) * | 2021-05-19 | 2021-09-17 | 中国航天科工集团第二研究院 | Human behavior electroencephalogram signal classification method and system |
| CN113870901A (en) * | 2021-09-26 | 2021-12-31 | 西北大学 | Voice emotion recognition method based on SVM-KNN |
| CN113988143A (en) * | 2021-12-27 | 2022-01-28 | 北京航空航天大学杭州创新研究院 | Bad track automatic identification method based on optical pump magnetometer measurement brain magnetism |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016162820A1 (en) * | 2015-04-07 | 2016-10-13 | Tata Consultancy Services Limited | System and method for estimating cognitive traits |
| CN106886792A (en) * | 2017-01-22 | 2017-06-23 | 北京工业大学 | A kind of brain electricity emotion identification method that Multiple Classifiers Combination Model Based is built based on layering |
| CN107220599A (en) * | 2017-05-16 | 2017-09-29 | 北京信息科技大学 | Image quality evaluating method based on EEG signal |
| WO2018014436A1 (en) * | 2016-07-18 | 2018-01-25 | 天津大学 | Emotion eeg recognition method providing emotion recognition model time robustness |
-
2020
- 2020-06-27 CN CN202010593365.5A patent/CN111832438B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016162820A1 (en) * | 2015-04-07 | 2016-10-13 | Tata Consultancy Services Limited | System and method for estimating cognitive traits |
| WO2018014436A1 (en) * | 2016-07-18 | 2018-01-25 | 天津大学 | Emotion eeg recognition method providing emotion recognition model time robustness |
| CN106886792A (en) * | 2017-01-22 | 2017-06-23 | 北京工业大学 | A kind of brain electricity emotion identification method that Multiple Classifiers Combination Model Based is built based on layering |
| CN107220599A (en) * | 2017-05-16 | 2017-09-29 | 北京信息科技大学 | Image quality evaluating method based on EEG signal |
Non-Patent Citations (2)
| Title |
|---|
| 赵亮;方芳;王伟;董寅冬;: "基于脑电信号间Granger因果关系特征的情感识别", 电子测量与仪器学报, no. 06 * |
| 郭金良;方芳;王伟;何寒娜;: "基于稀疏组lasso-granger因果关系特征的脑电信号情感识别", 模式识别与人工智能, no. 10 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112257658A (en) * | 2020-11-11 | 2021-01-22 | 微医云(杭州)控股有限公司 | Electroencephalogram signal processing method and device, electronic equipment and storage medium |
| CN112257658B (en) * | 2020-11-11 | 2023-10-10 | 微医云(杭州)控股有限公司 | Electroencephalogram signal processing method and device, electronic equipment and storage medium |
| CN112932502A (en) * | 2021-02-02 | 2021-06-11 | 杭州电子科技大学 | Electroencephalogram emotion recognition method combining mutual information channel selection and hybrid neural network |
| CN112932502B (en) * | 2021-02-02 | 2022-05-03 | 杭州电子科技大学 | Electroencephalogram emotion recognition method combining mutual information channel selection and hybrid neural network |
| CN113397567A (en) * | 2021-05-19 | 2021-09-17 | 中国航天科工集团第二研究院 | Human behavior electroencephalogram signal classification method and system |
| CN113870901A (en) * | 2021-09-26 | 2021-12-31 | 西北大学 | Voice emotion recognition method based on SVM-KNN |
| CN113870901B (en) * | 2021-09-26 | 2024-05-24 | 西北大学 | SVM-KNN-based voice emotion recognition method |
| CN113988143A (en) * | 2021-12-27 | 2022-01-28 | 北京航空航天大学杭州创新研究院 | Bad track automatic identification method based on optical pump magnetometer measurement brain magnetism |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111832438B (en) | 2024-02-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111832438A (en) | Electroencephalogram signal channel selection method and system for emotion recognition and application | |
| Park et al. | CNN-based learnable gammatone filterbank and equal-loudness normalization for environmental sound classification | |
| CN112244873A (en) | Electroencephalogram time-space feature learning and emotion classification method based on hybrid neural network | |
| Schmah et al. | Comparing classification methods for longitudinal fMRI studies | |
| CN112200016A (en) | Electroencephalogram signal emotion recognition based on ensemble learning method AdaBoost | |
| Hassan et al. | Epileptic seizure detection using a hybrid 1D CNN‐machine learning approach from EEG data | |
| Buza | Fusion methods for time-series classification | |
| CN112052874B (en) | Physiological data classification method and system based on generation countermeasure network | |
| Xie et al. | WT feature based emotion recognition from multi-channel physiological signals with decision fusion | |
| Ramzan et al. | Learning-based classification of valence emotion from electroencephalography | |
| Qian et al. | Geometric graphs from data to aid classification tasks with graph convolutional networks | |
| Foumani et al. | Disjoint-cnn for multivariate time series classification | |
| Shankar et al. | Classification of seizure types based on statistical variants and machine learning | |
| Abdulla et al. | An intelligent model involving multi-channels spectrum patterns based features for automatic sleep stage classification | |
| CN114676720A (en) | Psychological state identification method and system based on graph neural network | |
| Zhang et al. | An explainable and generalizable recurrent neural network approach for differentiating human brain states on EEG dataset | |
| CN113569924A (en) | Emotion identification classification method based on support vector machine multi-core cooperation | |
| Chatterjee et al. | Temporal window based feature extraction technique for motor-imagery EEG signal classification | |
| Pan et al. | A universal emotion recognition method based on feature priority evaluation and classifier reinforcement | |
| Trofimov et al. | A greedy feature selection algorithm for brain-computer interface classification committees | |
| Jiménez-Guarneros et al. | Cross-subject classification of cognitive loads using a recurrent-residual deep network | |
| CN112465054B (en) | FCN-based multivariate time series data classification method | |
| Ismail et al. | RL-ECGNet: resource-aware multi-class detection of arrhythmia through reinforcement learning | |
| CN114398991A (en) | Electroencephalogram emotion recognition method based on Transformer structure search | |
| CN113066544A (en) | FVEP characteristic point detection method based on CAA-Net and LightGBM |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |